AI spend management combines four controls: visibility (what was spent, on what), allocation (by team, repo, and person), caps (a limit that refuses a request rather than emailing you), and policy (which models are allowed at all). The market splits into cloud FinOps platforms that read invoices, LLM gateways that sit in the request path, observability tools that log requests, and built-in organization controls in the coding tools themselves. Only a control that sits in the request path can enforce a hard cap. Everything else is an alert after the money is gone.
- AI spend is usage-based, per-token, and mixed: prepaid subscriptions and metered API keys sit on the same invoice and behave in opposite ways.
- The four controls are visibility, allocation, caps, and policy. Most teams buy the first and skip the third.
- A cap is only hard if the vendor is in the request path. A BYOK developer with their own key cannot be capped, only observed.
- On a worked model, a 20-developer team at medium usage lands near $2,350 a month on Sonnet 5 rates. Model choice moves that by 2.5x before any other lever.
- The cheapest first move costs nothing: split the bill into prepaid and metered before you optimize anything.
Why AI spend is not SaaS spend
A SaaS line item is a seat count times a price, agreed in advance, changed at renewal. AI spend breaks every part of that sentence, and the tooling people reach for first was built for the sentence it breaks.
| Property | SaaS spend | AI spend | What it breaks |
|---|---|---|---|
| Unit | A seat, per month | A token, per request | Forecasting from headcount |
| Variance | Flat until renewal | 10x between two developers on the same team | Per-seat budgeting |
| Visibility | One invoice, one vendor | Subscriptions and metered keys mixed on the same team | Any single-number view |
| Latency to the number | Monthly, predictable | Real cost accrues in minutes, appears in reporting later | Reacting before the money is spent |
The mixed-rail problem is the one that catches finance teams. A developer on a Claude Max subscription and a developer holding an Anthropic API key can run identical work and produce completely different accounting: one is a prepaid capacity that costs nothing extra at the margin, the other is a metered bill that grows with every turn. Optimizing token usage on the prepaid rail saves you nothing in cash. Optimizing it on the metered rail is a direct refund. If your reporting cannot tell the two apart, every saving estimate you produce is wrong by an unknown amount.
The four controls that work
They are ordered. Each one is close to useless without the one before it, and each one is materially harder to install than the one before it.
01 · Visibility
One number per vendor per month, split into prepaid and metered, with the token counts behind it. This is table stakes and it is genuinely free: Anthropic ships a Usage and Cost Admin API at no charge, OpenAI exposes /v1/organization/costs, and every coding agent CLI writes a transcript to local disk that can be priced after the fact.
02 · Allocation
The same number, split by team, repo, project, and person. This is where most programs stall, because tokens carry no tag. You get attribution from whatever dimension the vendor happens to expose: an API key, a workspace, a project id, a seat, or the working directory recorded in the agent's own log.
03 · Caps
A limit that refuses the request. Not an email, not a dashboard threshold, not a Slack alert. The distinction matters more than any other line on this page, and it is the one vendors are vaguest about. See the next section.
04 · Policy
Which models are allowed, for whom, by default. This is the highest-leverage control and the least used, because it feels like it constrains engineers. It mostly constrains accidents: a default of Opus on work a mid-tier model handles is a 2.5x multiplier applied silently to everything. Writing that policy needs a defensible view of what each model is for, which is what our model directory is: one card per model, with the independently published coding scores next to the list price.
The build-versus-buy landscape
Four categories of product claim this space, and they solve genuinely different parts of it. Prices below were read from each vendor's own pricing page in August 2026; every one of them reprices, so check before you budget.
| Category | Examples | What it can actually enforce | Price |
|---|---|---|---|
| Provider-native reporting | Anthropic Usage & Cost Admin API, OpenAI /v1/organization/costs | Nothing. Reporting only, after the fact | Free |
| Cloud budget alerts | AWS Budgets | Alerts, plus scripted budget actions | Monitoring free; first two action-enabled budgets free, then $0.10 per budget per day |
| LLM gateway with budgets | Portkey, LiteLLM, Helicone | Hard per-key budgets and rate limits, because it is in the path | Portkey Production $49/mo for 100k logs, $9 per additional 100k; granular budget limits are Enterprise |
| LLM observability | Helicone, Langfuse, LangSmith | Cost per request, alerts, retention | Helicone Pro $79/mo; Langfuse Core $29/mo (100k units, then $8/100k); LangSmith Plus $39/seat |
| Built-in org controls | Continuum organizations, Cursor team spend limits | Model policy, weekly caps, member approvals | Continuum Team $25 per member per month |
The honest reading of that table: the free provider APIs give you control one, the gateways give you control three, and nothing gives you control two out of the box because allocation depends on how your organization is shaped rather than on which product you bought. A general FinOps platform such as Vantage or CloudZero is worth its price when AI is one line among many cloud vendors and finance needs it in the same report as everything else. It is overkill when AI coding tools are the whole problem.
Worked numbers for a 20-developer team
Every vendor comparison collapses into arithmetic eventually. Here is the arithmetic, with the assumptions stated so you can substitute your own. Rates are Claude Sonnet 5 at $2 per million input tokens, $10 per million output, and $0.20 per million cache-read tokens, checked against Anthropic's pricing page in August 2026.
| Profile | Cache reads / day | Fresh input / day | Output / day | Per day | Per month |
|---|---|---|---|---|---|
| Light | 2M | 0.3M | 60k | $1.60 | $33.60 |
| Medium | 8M | 1M | 200k | $5.60 | $117.60 |
| Heavy | 25M | 3M | 600k | $17.00 | $357.00 |
Twenty developers at the medium profile is $2,352 a month on metered API keys. The same twenty on Claude Team premium seats at $100 per seat on annual billing is $2,000 a month flat. At the heavy profile the metered bill is $7,140 and the seats are still $2,000. That is the whole plan-versus-API decision in three numbers, and it says the same thing every time: for full-time agent use, a subscription is a capacity purchase you will beat the marginal price on, and the reason to hold API keys is not cost.
A maturity model you can actually place yourself on
Five stages. Most engineering organizations in August 2026 are at stage 1, and the FinOps Foundation's 2026 survey found 98% of respondents now manage AI spend at all, up from 63% a year earlier. The practice is a year old in most companies, so being early is normal.
- Invisible. An invoice and a hunch. Nobody can say which team, repo, or model produced the number.
- Visible. One number per vendor per month, split prepaid versus metered. Achievable in an afternoon with free provider APIs.
- Allocated. The number splits by team, repo, and person, and the split is reproducible next month without a spreadsheet exercise.
- Capped. There is a ceiling that refuses work, and a documented path for someone to ask for more.
- Governed. Model allowlists, defaults, and an approval flow, reviewed on a cadence rather than after an incident.
The jump that pays for itself fastest is 1 to 2, because it is free and it usually surfaces one large obvious thing: an idle subscription, a runaway CI job holding a metered key, or a default model nobody chose. The jump that is hardest is 3 to 4, because it requires either that your team routes through something you control or that your vendor sells a real cap.
What not to do
- Do not buy a platform before you can answer the prepaid-versus-metered question. The answer changes which product category you need, and it is free to obtain.
- Do not set a cap you will always override. A ceiling that gets raised on every request teaches everyone that the ceiling is decorative, and the approval queue becomes pure friction with no control attached.
- Do not chase token efficiency on a prepaid rail. Trimming prompts on a subscription that is nowhere near its limit produces a slide, not a saving.
- Do not ban the expensive model outright. The failure mode is worse than the spend: work that needed the frontier model gets done badly three times instead of well once. Route it, default away from it, and require nothing.
- Do not treat per-seat pricing as a cost control. It is a cost cap and a usage blindfold at the same time. Flat seats hide the fact that three developers are producing most of the value and six are not using the tool at all.
Questions people ask
What is AI spend management?
AI spend management is the practice of making AI usage visible, attributable, and limited: knowing what was spent on which model by whom, allocating it to teams and projects, and enforcing ceilings before the spend happens. It differs from SaaS spend management because the billing unit is a token consumed per request rather than a seat priced per month, and because subscriptions and metered API keys usually sit on the same team.
How much does AI coding cost per developer per month?
On metered API keys at Claude Sonnet 5 rates ($2 in, $10 out, $0.20 cache reads per million tokens, checked August 2026), a modelled medium-usage developer lands near $118 a month and a heavy user near $357. Subscription tiers cap that: Claude Pro is $20 a month billed monthly, Max tiers are $100 and $200, and Claude Team premium seats are $100 a seat on annual billing. Full-time agent users almost always beat the metered price with a subscription.
What is the difference between an AI spend cap and an AI spend alert?
A cap refuses the request. An alert tells you the money is already gone. Only a system that sits in the request path, such as a gateway or a hosted inference endpoint, can implement a real cap. Tools that read a provider usage API after the fact can only alert, because that data arrives minutes to hours after the tokens were billed.
Do I need a FinOps platform for AI spend?
Only if AI is one line among many cloud vendors and finance needs a single report. If AI coding tools are the whole problem, the free provider APIs plus a gateway or built-in organization controls cover more of it for less money. Buy the platform for consolidation, not for AI-specific insight.
How do I track AI spend across multiple providers?
Three options, in increasing order of effort. Poll each provider's usage API and join the results yourself, which is free but per-vendor. Route everything through one gateway so there is a single log, which costs latency and a subscription. Or read the transcripts each agent CLI already writes to local disk and price them per model, which is retroactive and covers subscription usage the provider APIs report poorly.
Sources
Every figure above was read from these pages on August 2026. Vendors reprice without notice; if you find a stale number, tell us.